Chatbots are a gimmick, a hype, a solution to a problem that doesn’t exist, etc. All true, to a certain degree. After climbing out of the trough of disillusionment, we may yet see some productivity coming from our robot overlords.
I wouldn’t eat up your time with another article about the product/market fit of conversational UIs. Instead, let’s talk about the design effort that goes into building such an interface and how it boils down to building a genuine conversation.
UX design is a communication skill. Good design is based on exchanging messages one way or another. Be it through visuals, pixels in a browser, or chat bubbles in a messaging bot. The term “conversational interface” is somewhat redundant—every interface converses in some way.
However, what I find interesting about chatbots is why designers struggle to create really good ones. The http://chatbot.fail website gives a fine summary of what can go wrong when building a conversational interface.
We’ve been creating chatbots based on our years of experience in designing GUIs, websites, and apps. But notably, we never closed the loop. If we can identify the things we can learn from designing chatbots, we can generalize them into best practices in the wider field of UX design.
As an illustration, I discovered some remarkable insights while creating a chatbot for a Belgian bank. Those insights changed the way I look at UX challenges, and I believe they might be useful for any project—not just chatbots.
1. User feedback is instant, free, spontaneous
Whenever a user asks something to a chatbot, the interface should respond sensibly. If the algorithm is somehow not able to do that, a human takes over. The chatbot needs to be adapted in order to prevent a human intervention in the future.
This learning cycle is so blatantly obvious in chatbots that you wouldn’t dare to overlook it. Moving over to the more traditional UX projects, it can be a challenge to capture the same fidelity of user feedback in a website or a mobile app for example.
In user interviews, we often inquire for top tasks. In chatbots, we receive the unsuccessful top tasks neatly tagged in the bot’s back end (“human intervention,” “could not respond,” etc.).
Technology exists today to leverage our human interventions into automated self-improvement of the artificial intelligence. This lean build, measure, learn loop has never been shorter and we should strive to apply it to more traditional UX challenges as well. Ask yourself: How easily can a user experience correct itself, without design intervention?
2. How to AI
Artificial intelligence is so omnipresent in today’s literature that it would be fairly unlikely to design a user experience that isn’t influenced by this technology. And still, this project was the first real (not your average recommendation engine) AI-supported experience. Using Natural Language Understanding algorithms was a first when designing chatbots. AI is almost a prerequisite for designing a proper chatbot. The decision-tree logic doesn’t cut it and is not scalable given the variety of possible free-text user inputs.
Coming back to more traditional platforms, one could argue that this variety in user intent is generally the same. And still, we confront the user with a fixed set of navigation items, similar to any visitor. When designing interfaces, AI is there to help us augment experiences. To sweeten the deal: It does not require any introduction to our users—personalized feeds in social media are a commodity.
Your concerns that your engineering team doesn’t have any experience with emerging tech? Probably justified and yet another reason to add it to your next design proposal. If you start now, you’ll get there sooner or later. There is no “catch” in AI—only a chase that gets progressively interesting.
3. Not all content is dialogueable
When restricting a user experience to a single input field, you’ll find that some interactions require a lot of interaction cost. You wouldn’t want a chatbot to list a complete product catalogue. Conversational interfaces are a complementary interface.
Dialogue (text and voice) works for qualitative replies for concrete questions. Hands-on inquiries, transactional status updates, specific product information are all relevant use cases.
But when users required detailed instructions or tutorials for example, we opted for video instead of text. This guarantees a usable and fixed information flow, something we cannot guarantee when using chat. A fixed flow of information works better in this case, not filling the chat dialog with a wall of text.
It becomes extra challenging when adding voice into the mix. Some UX patterns simply don’t work without a visual representation. In a text-based interface a platform can still use links, attachments, images, and carousels to convey a message.
In voice you can’t. It’s impossible to spell out a hyperlink and expect users to memorize and navigate to it. However, oddly enough, some radio commercials do not seem to get this.
There is no way “out” of the interface. Audio is the only input/output you can expect. Restricting your skill to specific inquiries, status updates, and low-involvement transactions is key—at least until the capabilities of voice interfaces are augmented with other interfaces.
Related: Designing a conversational interface
In the context of this article, let’s take a look at the learnings we can implement in our everyday UX workflow.
Well, we often start with specific content. What I’m adding to my design process is a deep dive into what content types work best for the content I’m given.
The following questions come to mind:
- Given the platform we’re building, are we using the correct content type for this content?
- Should we communicate this content here? Is there a more appropriate place or platform to communicate this?
- Is the content sufficiently actionable for the limited affordance a platform gives us?
- How can we link or refer to other available content (including different content types) in a sensible way?
4. Who’s talking?
Building a chatbot starts with building a personality. The importance of this step cannot be overestimated. And it baffles me why I didn’t do it sooner for any other UX project. You can elaborate as far as you have to, but our personality described a gender, an age, an attitude description, typical filler vocabulary, and a personality background.
This makes sense for a chatbot, but equally so for any platform you’re building.
The way to differentiate no longer lies in the technology—it’s about how designers can use business and technology. Below, the excerpt from John Maeda explains it all (watch until the end).
Users’ expectations towards what digital platforms can offer are converging. And unless you’re really selling some disruptive technology, you’ll have more success differentiating your brand. This involves brand personality, tone of voice, brand values, visual identity and style, etc.
I know, I know—every brand document has this somewhere. But I’m sure I’m not the only one to never revisit this section after reading it. Whether you’re building a chatbot or not, you’ll find guidance and design inspiration in constraining yourself to a specific personality you’re impersonating in your communication. As I just mentioned: Design is communication, so isn’t it only logical that the one communicating also has a specific personality?
We talk so much about who we’re talking to that we may forget who is actually talking.
5. IA is AI’s best friend
Chatbots have an extremely flat information architecture—all information is accessible for a user’s question, and one might mistakenly think that there isn’t a need for a hierarchical structure.
Turns out that users still expect guidance, and they want to experience an information scent that gives them clues on where to start.
Moreover, you cannot start flooding complete answers to a single user’s intent. You have to throttle your information stream, and disclose it progressively during the conversation.
So don’t neglect your card-sorting skills. They have never been more relevant. Although AI may help in tailoring the user experience to a personal level, the algorithm still requires some information architecture to work with.
Ironically, machine learning algorithms exist to cluster unstructured data, but they often require a training set, which feeds the algorithm with something human. Apparently, it seems that AI can’t live without some user-oriented IA.